Latent symbolic circuit in large language models grows with more examples
In-Context Learning Amplifies a Latent Symbolic Circuit
Machine LearningArtificial IntelligenceComputation and Language
Summary
Large language models can learn new rules simply by seeing a few examples, but it’s unclear how their thinking parts wake up as they get more examples. The authors found a three-part reasoning process inside these models that starts working even before the model gets good at the task. This process gets stronger as the model sees more examples, and by tweaking parts of it, performance can improve a lot even without examples. So, the model’s internal setup for thinking abstractly is already there before it learns, and examples mainly help turn this setup on.
What this means in practice
- •For language model engineers: Improve few-shot model performance by targeting internal symbolic circuits for activation and control.
- •For ai service developers: Enhance prompt designs using function vectors to boost zero-shot and few-shot accuracy in deployed language applications.$Commercial implications: Enables more effective commercial AI assistants and chatbots that perform complex reasoning with fewer examples.
Authors
Melissa Wessel
Abstract
Large language models can learn abstract rules from just a few in-context examples, but how their internal mechanisms activate as examples accumulate is not well understood. We trace a three-stage symbolic reasoning circuit (abstraction, induction, retrieval) across shot counts in three model families and find it is detectable and functional well before the model achieves high accuracy. Per-head causal contribution grows up to 8x from 1- to 10-shot, and cross-shot activation patching raises accuracy from 1% to 56% at 0-shot and 17% to 88% at 1-shot. Function vectors scaled and injected at 0-shot rescue accuracy up to 86%, largely substituting for the induction stage but depending critically on an intact downstream retrieval stage. The infrastructure for abstract rule-following is present in the weights before any demonstrations; in-context examples, function vectors, and related interventions appear to supply input to the same latent circuit.